The Agent Skills Directory

[PROMPT_INJECTION]: No instructions attempting to bypass safety filters or override system prompts were detected. The instructions are focused on technical implementation details for GPU kernels.
[DATA_EXPOSURE]: No hardcoded credentials or access to sensitive system file paths (like SSH keys or AWS configs) were found. The skill references project-specific files such as devices.json and flash-mla.md which are relevant to the kernel development context.
[OBFUSCATION]: No obfuscated code, zero-width characters, homoglyphs, or encoded strings (Base64/Hex) intended to hide malicious intent were identified.
[EXTERNAL_DOWNLOADS]: The skill does not contain any commands to download or install external packages from untrusted sources. It mentions standard libraries like torch and domain-specific tools like cuda-tile-translate and cuda.tile_experimental in a descriptive context.
[COMMAND_EXECUTION]: The provided Python code snippets are templates for defining GPU kernels and do not include dangerous shell command executions or arbitrary code evaluation patterns.
[DYNAMIC_EXECUTION]: The skill describes a DSL that involves JIT compilation (compile_tile) and autotuning for GPU kernels. This is standard functionality for performance optimization in machine learning infrastructure and does not involve processing untrusted input in an unsafe manner.
[DYNAMIC_CONTEXT_INJECTION]: The !command syntax for load-time execution was not found in the SKILL.md file.

design-cutile-dsl-kernel